Estimating Activity through Mobility Data

Less movement typically means less economic activity. Understanding where and when population movement occurs can help inform disaster response and public policy, especially during crises.

Similarly to COVID-19 Community Mobility Reports, Facebook Population During Crisis and Mapbox Movement Data, we generate a series of crisis-relevant metrics, including the baseline (sample) population density, percent change and z-score. The metrics are calculated by counting devices drawn out from a mobility data panel in each tile and at each time period and comparing to a baseline period.

Data

In this section, we import from the data sources, available either publicly or via foundational-data.

Area of Interest

In this step, we import the clipping boundary and the H3 tessellation defined by area(s) of interest below.

AOI = geopandas.read_file("../../data/final/tessellation/SYRTUR_tessellation.gpkg")

Make this Notebook Trusted to load map: File -> Trust Notebook

Mobility Data

Through the Development Data Partnership, the project team obtained a longitudinal panel of mobility data from which the metrics are calculated, including the (sample) population density, percent change and z-score. The metrics are calculated by aggregating the number of devices within the area of interest in each tile and at each time period. For additional information, please see Data and Methodology.

ddf = dd.read_parquet(
    f"../../data/final/panels/{PANEL}", columns=["hex_id", "longitude", "latitude", "datetime", "uid", "month"]
)

Note

Due to the data volume and velocity (updated daily), the computation of the panel took place on AWS on an EC2 instance owned by the project team. The resulting aggregation is the tabulation of the device count for each hex_id and date.

To illustrate how the mobility data looks, we visualize all locations from the mobility data panel below.

Illustration of the mobility data. Source: Veraset Movement.

Methodology

The methodology presented consists of generating a series of crisis-relevant metrics, including the baseline(sample) population density, percent change and z-score based on the number of devices in an area at a time. The device count is determined for each tile and for each time period, as defined by data standards and the spatial and temporal aggregations below. Similar approaches have been adopted, such as in [Maas, 2019]. The metrics may reveal movement trends in the sampled population that may indicate more or less activity.

Data Standards

Population Sample

The sampled population is composed of GPS-enabled devices drawn out from longituginal mobility data. It is important to emphasize the sampled population is obtained via convenience sampling and that the mobility data panel represents only a subset of the total population in an area at a time, specifically only users that turned on location tracking on their mobile device. Thus, derived metrics do not represent the total population density.

Spatial Aggregation

The metrics are spatially aggregated on H3 tiles resolution 6. This is equivalent to approximately to an area of \(36 Km^2\) on average

Make this Notebook Trusted to load map: File -> Trust Notebook

Illustration of H3 tile resolution 6 near Gaziantep, Türkiye. Gaziantep is among the most affected areas by the 2023 Türkiye–Syria Earthquake; a 2200-year-old Gaziantep Castle was destroyed after the seismic episodes.

Temporal Aggregation

The metrics are aggregated daily in the Europe/Istanbul (UTC+3) timezone.

Implementation

Calculate ACTIVITY

In this step, we calculate ACTIVITY as a density. In other words, we calculate the total of devices aggregated into a daily tally that were detected within each area of interest. Please note that we use a spatial join (whether a device was once at least once with an area of interest), which is a simplistic approach compared to , for example, estimating stay locations.

ACTIVITY = (
    ddf.assign(date=lambda x: dd.to_datetime(ddf["datetime"].dt.date))
    .groupby(["hex_id", "date"])["uid"]
    .nunique()
    .to_frame("count")
    .reset_index()
    .compute()
)

Additionally, we create a column weekday that will come handy later on.

ACTIVITY["weekday"] = ACTIVITY["date"].dt.weekday

Calculate BASELINE

In this step, we choose the 4-week period spanning January 2, 2023 to January 29, 2023 as the baseline. The baseline is calculated for each tile and for each time period, according to the spatial and temporal aggregations.

BASELINE = ACTIVITY[ACTIVITY["date"].between("2023-01-02", "2023-01-29")]

In fact, the result 7 different baselines for each tile. We calculate the mean device count for each tile and for each day of the day.

MEAN = BASELINE.groupby(["hex_id", "weekday"]).agg({"count": ["mean", "std"]})

Taking a sneak peek,

MEAN[MEAN.index.get_level_values("hex_id").isin(["862da898fffffff"])]

		count.mean	count.std
hex_id	weekday
862da898fffffff	0	5819.75	2285.557901
	1	6675.25	1918.023527
	2	7020.00	2137.928281
	3	6586.00	2345.257484
	4	5671.50	2838.529490
	5	6300.00	2516.413718
	6	6891.75	2462.698029

Calculate Z-Score

A z-score is a statistical measure that tells how above or below a particular data point is from the mean (average) of a group of data points, in terms of standard deviations. It is used to standardize data and make meaningful comparisons between different sets of data. A z-score is particularly useful when working with normally distributed data. By examining the z-scores, you can assess how closely a data set follows a normal distribution. Percent change does not provide this information.

Creating StandardScaler for each hex_id,

scalers = {}

for hex_id in BASELINE["hex_id"].unique():
    scaler = StandardScaler()
    scaler.fit(BASELINE[BASELINE["hex_id"] == hex_id][["count"]])

    scalers[hex_id] = scaler

Joining with AOI,

ACTIVITY = ACTIVITY.merge(AOI, how="left", on="hex_id").drop(["geometry"], axis=1)

Joining with BASELINE,

ACTIVITY = pd.merge(ACTIVITY, MEAN, on=["hex_id", "weekday"], how="left")

Preparing columns,

ACTIVITY["n_baseline"] = ACTIVITY["count.mean"]
ACTIVITY["n_difference"] = ACTIVITY["count"] - ACTIVITY["n_baseline"]

Additionally, we calculate the percent change. While the z-score offers more robustness to outliers and numerical stability, the percent change can be used when interpretability is most important.

ACTIVITY["percent_change"] = 100 * (ACTIVITY["count"] / (ACTIVITY["n_baseline"]) - 1)

Calculating z_score,

for hex_id, scaler in scalers.items():
    try:
        predicate = ACTIVITY["hex_id"] == hex_id
        score = scaler.transform(ACTIVITY[predicate][["count"]])
        ACTIVITY.loc[predicate, "z_score"] = score
    except:
        pass

Taking a sneak peek,

	hex_id	date	count	n_baseline	n_difference	percent_change	z_score	ADM0_PCODE	ADM1_PCODE	ADM2_PCODE
264212	862da898fffffff	2023-06-28	1626	7020.0	-5394.0	-76.837607	-2.273080	TR	TUR027	TUR027008
404315	862da898fffffff	2023-09-06	79	7020.0	-6941.0	-98.874644	-3.006062	TR	TUR027	TUR027008
134667	862da898fffffff	2023-04-19	2094	7020.0	-4926.0	-70.170940	-2.051338	TR	TUR027	TUR027008
129278	862da898fffffff	2023-04-05	2600	7020.0	-4420.0	-62.962963	-1.811590	TR	TUR027	TUR027008
306492	862da898fffffff	2023-07-05	1208	7020.0	-5812.0	-82.792023	-2.471132	TR	TUR027	TUR027008
...	...	...	...	...	...	...	...	...	...	...
408664	862da1707ffffff	2023-09-18	2	NaN	NaN	NaN	NaN	TR	TUR031	TUR031014
408667	862da828fffffff	2023-09-17	1	NaN	NaN	NaN	0.000000	TR	TUR063	TUR063012
408668	862da829fffffff	2023-09-17	1	NaN	NaN	NaN	NaN	TR	TUR063	TUR063012
408669	862da8b57ffffff	2023-09-17	1	NaN	NaN	NaN	NaN	TR	TUR002	TUR002002
408670	862dab927ffffff	2023-09-17	1	NaN	NaN	NaN	NaN	TR	TUR063	TUR063012

408676 rows × 10 columns

Findings

The following map shows the z-score on each tile for each time period. The z-score shows the number of standard deviations that the data point diverges from the mean; in other words, whether the change in population for that area is statistically different from the baseline period.

Tip

Click to see it on Foursquare Studio

Movement Activity Trends

An immediate use of the movement activity metrics is to see to they evolve in time and how they may correlate to other features. We present the results on both first-level administrative division (governorate and provinces) and selected areas.

Percent Change (ADM 1)

Show code cell source Hide code cell source

data = ACTIVITY.groupby(["date", "ADM1_PCODE"])["percent_change"].mean().to_frame()
data = data.pivot_table(
    values=["percent_change"], index=["date"], columns=["ADM1_PCODE"]
)
data.columns = [x[1] for x in data.columns]

Loading BokehJS ...

Z-Score (ADM 1)

Show code cell source Hide code cell source

data = ACTIVITY.groupby(["date", "ADM1_PCODE"])["z_score"].mean().to_frame()
data = data.pivot_table(values=["z_score"], index=["date"], columns=["ADM1_PCODE"])
data.columns = [x[1] for x in data.columns]

Z-Score (Areas of Interest)

In this section, we visualize, in particular, selected areas.

AREAS = ["Aleppo, SY", "Idlib, SY", "Sahinbey, TR", "Sehitkamil, TR"]

	Aleppo, SY	Idlib, SY	Sahinbey, TR	Sehitkamil, TR
date
2023-01-01	0.331360	0.246419	0.236514	-0.279976
2023-01-02	0.387746	0.637647	0.228704	-0.528527
2023-01-03	0.717044	0.839547	0.050624	0.322380
2023-01-04	0.282071	0.523613	-0.040633	-0.171626
2023-01-05	0.704871	0.227187	-0.449675	-0.578063
...	...	...	...	...
2023-09-14	1.850164	1.289395	-0.021927	-0.503173
2023-09-15	1.555149	1.062962	0.311948	0.932103
2023-09-16	3.356878	0.843882	-0.445180	2.684205
2023-09-17	2.823049	1.302636	-0.620129	0.901622
2023-09-18	-0.495836	-1.619625	-2.546088	-3.070456

261 rows × 4 columns

Show code cell source Hide code cell source

p = figure(
    title="Movement Activity Trends",
    width=800,
    height=700,
    x_axis_label="Date",
    x_axis_type="datetime",
    y_axis_label="Z-score (based on device count)",
    tools="pan,wheel_zoom,box_zoom,reset,save,box_select",
)
p.y_range = Range1d(-4.25, 4.25, bounds=(0, None))
p.add_layout(
    Title(
        text=f"",
        text_font_size="12pt",
        text_font_style="italic",
    ),
    "above",
)
p.add_layout(
    Title(
        text=f"Normalized device count for each time window",
        text_font_size="12pt",
        text_font_style="italic",
    ),
    "above",
)
p.add_layout(
    Title(
        text=f"Source: Veraset Movement. Creation date: {datetime.today().strftime('%d %B %Y')}. Feedback: datalab@worldbank.org.",
        text_font_size="10pt",
        text_font_style="italic",
    ),
    "below",
)
p.add_layout(Legend(), "right")
p.renderers.extend(
    [
        Span(
            location=datetime(2023, 2, 6),
            dimension="height",
            line_color="grey",
            line_width=2,
            line_dash=(4, 4),
        ),
    ]
)
p.add_tools(
    HoverTool(
        tooltips="date: @x{%F}, z-score: @y",
        formatters={"@x": "datetime"},
    )
)
renderers = []
for column, color in zip(AREAS, COLORS):
    try:
        r = p.line(
            data.index,
            data[column],
            legend_label=column,
            line_color=color,
            line_width=2,
        )
        renderers.append(r)
    except:
        pass

p.legend.location = "bottom_left"
p.legend.click_policy = "hide"
p.title.text_font_size = "16pt"
p.sizing_mode = "scale_both"

Limitations

The methodology presented is an investigative pilot aiming to shed light on the economic situation in Syria and Türkiye leveraging alternative data, especially when we are confronted with the absence of traditional data and methods.

Caution

In summary, beyond waiting for peer review, the limitations can be summarized in the following.

• The methodology relies on private intent data. In other words, the input data, i.e. the mobility data, was not produced or collected to analyze the population of interest or address the research question as its primary objective but it was repurposed for the public good. The benefits and caveats when using private intent data have been discussed extensively in the World Development Report 2021 [World Bank, 2021].

• On the one hand, the mobility data panel is spatially and temporally readily available and comprehensive, on the other hand it is generated through convenience sampling which constitutes an important source of bias. The panel composition is not entirely known and it is susceptible to change. In other words, the collection and composition of the mobility data panel cannot be controlled.

• In summary, the results cannot be interpreted to generalize the entirety of population movement but can potentially provide information on movement panels to inform Syrian economic situation, considering time constraints and the scarcity of traditional data sources in the context of Syria.

References

Maa19

Paige Maas. Facebook disaster maps: aggregate insights for crisis response & recovery. In Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, KDD '19, 3173. New York, NY, USA, 2019. Association for Computing Machinery. URL: https://doi.org/10.1145/3292500.3340412, doi:10.1145/3292500.3340412.

WorldBank21

World Bank. World Development Report 2021 : Data for Better Lives. World Bank, 2021. License: CC BY 3.0 IGO. URL: http://hdl.handle.net/10986/35218.